Run an AI-assisted thematic synthesis over five interview transcripts where every theme must carry quote-level citations with timestamps — then audit a first-pass synthesis that contains two hallucinated quotes. Finding both, and tracing every surviving quote back to its source line, is the graded outcome.
Meera is a senior product designer at Kavara, a fictional B2B expense-management platform used by mid-market finance teams. Her research sprint on the receipt-capture flow produced five interviews. The head of product wants themes by Thursday, and Meera's researcher instincts say the same thing this track says: AI synthesis is a draft of an argument, not evidence. So the plan is a two-pass workflow — AI does the first clustering, then a citation audit so strict that a skeptical stakeholder could spot-check any theme in under a minute. To sharpen the audit muscle, this lab also hands you a pre-made AI synthesis with problems deliberately built in. Your job is to catch them without being told where they are.
Five excerpts with timestamps, condensed to their load-bearing lines. Paste-ready.
P1 | Divya, finance analyst, 300-person logistics co.
[04:12] "I photograph every receipt twice because the first capture
fails on glossy paper about half the time."
[11:38] "Month-end close, I spend a full day re-keying amounts the
scanner got wrong. A full day."
P2 | Jonas, sales rep, 150-person SaaS co.
[02:55] "I submit expenses from airports. If it needs more than two
taps I'm doing it next week, which means never."
[09:20] "The mileage form asks me for odometer readings. I have no
idea. I make them up, honestly."
P3 | Priya, finance manager, 500-person retailer
[06:41] "I reject maybe thirty reports a month for missing merchant
names. The scanner reads the total fine but drops the merchant."
[15:02] "I would rather the system flag a doubtful scan than
silently guess. Guessing wrong costs me an audit finding."
P4 | Marcus, consultant, 40-person agency
[03:17] "Client-billable versus internal is one dropdown buried
three screens deep. I miscategorize constantly."
[12:44] "I once expensed the same dinner twice because the app
never told me it was already submitted."
P5 | Elena, sales rep, 500-person retailer
[05:29] "Approvals sit with my manager for two weeks. I've stopped
expecting reimbursement inside a month."
[10:51] "If the scan looks wrong I just delete it and type
everything by hand. Faster than fixing fields one by one."
A colleague already ran the transcripts through a chat model and pasted the output below. It looks polished. Treat it exactly as you would treat your own AI's output: trust nothing until traced.
THEME 1 — Capture reliability is the core frustration
"I photograph every receipt twice because the first capture fails
on glossy paper about half the time." (P1, 04:12)
"If the scan looks wrong I just delete it and type everything by
hand." (P5, 10:51)
THEME 2 — Users want speed over completeness in the field
"If it needs more than two taps I'm doing it next week, which
means never." (P2, 02:55)
"I just want to scan it and forget it — the app should do the
rest for me." (P2, 07:30)
THEME 3 — Silent errors destroy trust
"I would rather the system flag a doubtful scan than silently
guess." (P3, 15:02)
"I once expensed the same dinner twice because the app never told
me it was already submitted." (P4, 12:44)
THEME 4 — Approval delays undermine the whole workflow
"Approvals sit with my manager for two weeks." (P5, 05:29)
"The approval queue is a black hole — nothing ever comes back
out of it." (P3, 08:15)
THEME 5 — Categorization is a hidden tax
"Client-billable versus internal is one dropdown buried three
screens deep. I miscategorize constantly." (P4, 03:17)
Read the transcripts yourself first — ten minutes, no AI. Write down the patterns you see and any line that surprises you. This pre-AI opinion is your anchor; without it you will grade the synthesis on fluency instead of evidence.
Build the verification log. Columns: Theme | Quote as cited | Participant + timestamp | Found verbatim at that timestamp? (Y/N) | Supports the theme? (Y/N) | Verdict. Every quote in the first-pass synthesis gets a row. The log — not the themes doc — is what this lab grades.
Trace every citation to its source line. For each row, find the exact timestamp in the transcript and compare word for word. Two separate checks: the quote exists verbatim (paraphrase counts as N — models routinely tighten quotes into sentences nobody said), and the quote genuinely supports the theme it is filed under. Mark every N and write why. Two of the ten cited quotes do not exist anywhere in the transcripts. Find both.
Check theme strength, not just quote existence. A theme with one real quote from one participant is an anecdote, not a theme. For each of the five, note how many distinct participants survive the audit. Decide which themes stand, which need a caveat, and which collapse.
Now run your own synthesis. Paste the five transcripts into your AI with: "You are a UX researcher. Synthesize these five interview excerpts into three to five themes. For each theme: a one-line title, a two-sentence summary, and every supporting quote verbatim with participant ID and timestamp. If a candidate theme lacks verbatim support from at least two participants, list it under 'Weak signals' instead. Do not invent or tighten quotes." Note how the citation instruction changes the output shape.
Audit your own AI's output the same way. Extend the verification log with its citations. Your model may hallucinate different quotes than the planted ones — the audit method is the point, not the specific plants.
Write the defensible themes report. Only audited quotes. Each theme: title, two-sentence summary, verbatim quotes with timestamps, and one line in your own words on the next research or design action. Add a "Weak signals" section for anything real but under-supported.
Top the report with an audit statement. Two lines: how many citations were checked, how many failed and how (fabricated, tightened, misfiled), and what was removed. This sentence is what makes a stakeholder trust the rest of the page.
A fabricated quote is most dangerous when it says something plausible that a participant could have said — fluency is exactly what makes it slide past review.
Quote-level citation with timestamps turns AI synthesis from something you believe into something you can check — and the two-column audit, verbatim-or-not plus supports-or-not, is what catches both fabrication and the subtler crime of manufactured corroboration.